Beyond Replace or Retrain: The Worker Agent

The wrong order

Late last month, Gartner told employers something many of their workers already suspected. By 2029, the research firm projects, roughly 30 percent of the employees who were let go and replaced by AI will be hired back, usually at a higher cost than it took to lose them. Its HR analyst put the diagnosis in one line: many organizations are making their AI and workforce decisions in the wrong order.

The order she means runs like this. Deploy the systems. Let them learn from the people already doing the job. Cut the people. Discover what the systems cannot do. Rehire, at a premium, some fraction of the expertise that walked out the door. The pattern now has a face: Amazon, after roughly 30,000 corporate layoffs, has been emailing former employees, including some it laid off, about open AI and cloud roles under what one recruiter called a “boomerang reengagement initiative.” Amazon says rehiring alumni is routine. Even taken at face value, the point stands. The expertise was treated as surplus, then discovered to be scarce.

So the public conversation has settled into a two-option frame. Either companies replace workers with AI, or they retrain workers to use AI. Replace or retrain. Both options share an assumption worth noticing: the AI belongs to the employer. Even the best version of retrain shows it. IKEA’s parent, rather than cut the 8,500 people its chatbot could have displaced, retrained them as interior design consultants. But the AI is still IKEA’s. The worker was moved out of its way, not equipped to meet it.

Neither option gives the worker an AI of her own. That is the gap this post is about. It turns out this is the same human-side gap I wrote about last week — for learners generally, and in particular those students and parents facing Google’s bundling of Search and AI into their classrooms. There I proposed moving beyond the binary of ban or bundle. Here, the choice on offer is to replace employees or to retrain them. I believe there is a third option worth considering: the worker brings her own agent.

What the employer’s AI takes

To see why that frame is wrong, look at what the replacement path actually consumes. If you are the employee, it consumes you.

Everything you do at work digitally leaves a wake: every email, chat message, document edit, prompt, and correction. Employer systems no longer capture only what you have produced voluntarily. They increasingly capture what you know and how you think: your reasoning patterns, your judgment, the way you frame a problem. We call this cognition capture, and it is no longer speculative. A TIME investigation this spring described Oracle employees who were instructed to train AI systems on their own work, and then laid off. The “digital twin of an employee” is now a named enterprise product category. And this past week, Fast Company reported on “workforce management” tools that help decide who is laid off, fed by keyboard activity, sick days, and “flight risk” scores. Workers have learned how those systems work, when they have learned at all, largely through lawsuits.

On factory floors the same thing is happening to the human body. Tesla is reported to be fitting workers with camera rigs so its robot can learn their motions. The through-line is straightforward. The binding constraint on replacing a worker is data about how that worker works. The employer’s AI learns from her, models her, and increasingly decides about her. And she has nothing comparable on her side of the interface.

This is also a breach of the oldest bargain in employment. A worker sells her present labor — the meetings attended, the documents written, the tasks performed. She has never sold a reproducible template of herself that the company can then keep forever. Work-for-hire gives the employer the outputs it paid for, not the producer. The law already binds the worker to the firm with duties of loyalty and confidentiality. No such duty runs the other way. The firm’s essence is protected against the worker. The worker’s essence is not protected against the firm.

Bring Your Own Bot

In prior blogs I described what GliaNet calls a Personal Augmentation Agent (PAA). This is a type of AI agent created and trained to represent the human side of the interface. A Worker Agent, then, is a PAA for the job. It works for the employee, and only the employee, inside a workplace where the employer already runs AI systems of its own. It is not a productivity feature the company hands out. It is the worker’s agent, chosen by her, loyal to her, and portable when she moves on.

In daily use, such an agent does what a good advisor does. It keeps her work-related data under her control. It coaches her on the skills the job demands, and those for the next job. It prepares her for the review, the pay negotiation, the scheduling dispute, with the facts organized and the options laid out. It helps her read the algorithmic systems that rate her, including the ones that decide who stays. And it governs, on her instructions, which of her knowledge and working methods flow to employer systems, under what terms, and for how long. When a system starts modeling her individually, her agent notices, documents it, and asks for the consent the rules require.

What makes the Worker Agent trustworthy is not the model inside it, but the duties around it. A Worker Agent is operated by what we call a Net Fiduciary™, an entity bound by charter to three enforceable duties to the person it serves: care (protect her from harm), fidelity (carry out her instructions with no conflicts), and loyalty (affirmatively promote her interests). The middle duty is the one that matters most at work, because the second master is always in the room. An agent that owes anything to the employer, or to the vendor that built it, is not a Worker Agent. It is an Employer Bot, wearing a friendly face.

The policy that lets this agent into the building is what we have termed “BYOB” — Bring Your Own Bot. Employers have done this before. When workers brought their own phones and laptops to the office, IT departments did not ban them. Instead, they built access controls that worked without owning the hardware. That same approach can work for agents. What is missing is not infrastructure. It is recognition that the worker has a right to an agent of her own.

Consider the moment it matters most: departure. A senior support engineer resigns after six years. Her employer spent two of them building an internal “expert assistant” trained on her tickets and escalation decisions, a fact her agent detected and documented at the time. At exit, her agent produces the inventory of models trained on her work. Her knowledge base returns to her own data store. The employer keeps the aggregate patterns it legitimately owns. A covenant confirms that no system will be run to answer “what would she do?” without her say (and then, perhaps, for a fee). The employee walks out with what she walked in with, plus six years of growth. Her own.

None of this is hypothetical for us. In October, GliaNet filed a funding application to the Humanity AI Open Call, together with the National Hispanic Health Foundation (NHHF). Our joint proposal is to build and pilot a Worker Agent for Hispanic nurses and health professionals. These professionals already work beside ambient scribes that record every patient encounter, coding engines that draft the bill, and productivity dashboards that score the shift. The design we have proposed is the one described here, with one feature worth underlining: the community writes the rules. A Steering Committee of NHHF fellows and members sets the Agent’s requirements, controls the data rules, selects the builder, and holds a veto over any feature or data flow it judges against members’ interests.

In short, GliaNet supplies the fiduciary charter and the governance, while the nurses supply the standards for what a loyal agent in their workplace must and must not do. That co-design process is the part of the model that travels. We stand ready to run this approach with any worker organization, union, professional association, or trade body whose members want an agent on their side. Our job would be to help them co-design and launch Worker Agents that do the kinds of things described in this post.

The right order

If the wrong order is capture, replace, regret, rehire, the right order is simpler: augment first. For employers it is also the cheaper order. A company that rehires 30 percent of its AI-replaced staff at a premium has paid three times: to lose the people, to discover the gap, and to bring them back.

We will not pretend that the business case will carry the day on its own. Some employers, weighing the same facts, will choose replacement, because on a spreadsheet it pencils out. That is why Bring Your Own Bot should be a matter of rights and counter-architecture, not a plea for enlightened self-interest. The State of California took the first steps this fall, restricting AI surveillance that infers workers’ emotions or collects their neural data, and barring sole reliance on automated systems to discipline or fire them. Those laws constrain the employer’s AI. The other half of the answer is to equip the worker with AI that answers to her.

If you are a worker: Ask what AI tools your employer deploys, what they model, and whether any system is trained on individual employees’ work. Keep your own record of your expertise outside company systems. Talk to colleagues; the law protects that conversation.

If you are an employer: Inventory every system that models individual workers, disclose it, and pilot a BYOB program with willing staff. Measure it against Gartner’s rehire premium.

If you make policy: Extend what Hollywood performers won for their faces and voices — consent, compensation, and control over replicas — to every worker’s cognition and skills. And protect the worker’s right to bring her own agent.

Replace or retrain is a choice between two ways of handing the worker’s future to the employer’s AI. The third option gives the worker an AI of her own. The companies now rehiring the people they replaced have learned, expensively, half of that lesson. The other half is that the worker should never have been on her own in the first place.

Further reading

• Richard Whitt, Beyond an AI Ban or Bundle: The Learner Agent (The Deeper Edge Blog, November 6, 2026).
• Richard Whitt, Why We All Need and Deserve Personal Augmentation Agents (GliaNet Concept Note), October 16, 2026.
• Gartner, via ITPro (Nov. 2, 2026) | Inc. on Amazon’s boomerang hiring | Fast Company, “How likely are you to be fired by AI?” (Nov. 4, 2026) | TIME on Oracle (Apr. 30, 2026).

Want to take this further?

To explore co-designing a Worker Agent with your union, professional association, or worker organization, get in touch.

Contact Us